Source-linked AI summary
CondLaneNet: a Top-to-down Lane Detection Framework Based on Conditional Convolution
Lizhe Liu, Xiaohao Chen, Siyu Zhu, Ping Tan
TL;DR
Lane detection methods still struggle with instance-level discrimination and complex lane topologies such as dense and fork lines. CondLaneNet detects lane instances before dynamically predicting each line’s shape using conditional convolution, row-wise formulation, and RIM. It achieves state-of-the-art performance across three benchmarks while combining accuracy with real-time efficiency.
Problem
Lane detection remains challenged by assigning predicted points to distinct instances and perceiving complex topologies such as dense and fork lines.
Method
CondLaneNet detects lane instances first, then dynamically predicts each instance’s shape through conditional convolution, row-wise formulation, and the Recurrent Instance Module.
Results
CondLaneNet refreshes state-of-the-art performance on CULane, CurveLanes, and TuSimple; its small version reaches 78.14 F1 at 220 FPS on CULane.
Takeaways & Limitations
The framework combines improved lane-instance discrimination and complex-topology handling with real-time efficiency on the evaluated benchmarks.
Abstract
from arXiv · showhide
Modern deep-learning-based lane detection methods are successful in most scenarios but struggling for lane lines with complex topologies. In this work, we propose CondLaneNet, a novel top-to-down lane detection framework that detects the lane instances first and then dynamically predicts the line shape for each instance. Aiming to resolve lane instance-level discrimination problem, we introduce a conditional lane detection strategy based on conditional convolution and row-wise formulation. Further, we design the Recurrent Instance Module(RIM) to overcome the problem of detecting lane lines with complex topologies such as dense lines and fork lines. Benefit from the end-to-end pipeline which requires little post-process, our method has real-time efficiency. We extensively evaluate our method on three benchmarks of lane detection. Results show that our method achieves state-of-the-art performance on all three benchmark datasets. Moreover, our method has the coexistence of accuracy and efficiency, e.g. a 78.14 F1 score and 220 FPS on CULane. Our code is available at https://github.com/aliyun/conditional-lane-detection.
1. Introduction
Lane detection remains challenged by instance-level discrimination and complex topologies such as dense and fork lines. CondLaneNet addresses these issues with conditional lane detection, row-wise shape formulation, and the RIM, achieving accurate and efficient detection.
- Instance-level discrimination remains difficult because methods must assign predicted lane points to different instances, while fixed lane classes limit detections to a predefined number.
- Complex topologies, including dense and fork lines, remain challenging, with few real-time perception studies addressing them.
- CondLaneNet combines conditional lane detection with a row-wise formulation to improve lane instance-level discrimination.
- The Recurrent Instance Module addresses lane detection for complex topologies such as dense and fork lines.
- 86.10 F1 on CurveLanes and 79.48 F1 on CULane establish state-of-the-art performance, while the small model reaches 78.14 F1 at 220 FPS on CULane.
2. Related Work
Prior lane detection methods use segmentation, anchors, row-wise prediction, or parametric curves, each addressing shape representation and efficiency differently. CondLaneNet builds on top-to-down detection and row-wise shape prediction while adding conditional framework components.
- Segmentation-based methods: Segmentation-based methods commonly require lane instance discrimination, using multi-class labeling or post-clustering strategies to separate lanes.
- Anchor-based methods: Anchor-based methods use top-to-down pipelines and relative-coordinate regression, with predefined anchors reducing the impact of limited visual cues.
- Row-wise detection methods: Row-wise methods predict lane locations for each row and exploit continuity, but instance-level discrimination remains their main problem.
- CondLaneNet framework: CondLaneNet detects proposal points at line starts, predicts dynamic convolution kernels, and uses a conditional shape head for row-wise lane descriptions.
- Parametric prediction methods: Parametric methods directly regress curve equations and can be highly efficient, but have not surpassed other methods in accuracy.
3. Methods
CondLaneNet detects lane instances from start-point proposals, then predicts each line’s shape with conditional convolution and a row-wise representation. Its Recurrent Instance Module supports multiple lane instances at one proposal point, addressing dense and forked topologies.
- 3. Methods: CondLaneNet represents each lane as an ordered coordinate set and structures the method around conditional lane detection, RIM, and framework design.The framework predicts a collection of lanes from an input image and organizes processing into instance detection, shape prediction, recurrent instance handling, and architecture components.
- 3.1. Conditional Lane Detection: Conditional lane detection first detects instances and then uses dynamic kernel parameters for instance-wise shape prediction.Unlike mask-based conditional instance segmentation, the method conditions shape prediction on dynamic kernels and uses a lane-specific representation.
- 3.1.1 Shape Prediction: The row-wise shape head predicts location, vertical range, and offsets, then aggregates row locations from bottom to top into lane lines.Expected abscissas provide sub-grid row locations, while the vertical range identifies traversed rows and the offset map refines horizontal positions.
- 3.1.2 Instance Detection: The proposal head detects lane instances using start points because their visual characteristics are clearer than those of slender, curved lane centers.A proposal heatmap is trained with focal loss, and a parameter map supplies dynamic kernel parameters for each detected proposal.
- 3.2. Recurrent Instance Module: RIM recurrently predicts continue-or-stop states and lane-wise kernel parameters, allowing one proposal point to guide multiple lane instances.This design addresses cases where fork lanes share a proposal point and supports subsequent instance-wise dynamic convolution.
4. Experiments
CondLaneNet is evaluated on three lane-detection benchmarks using standardized metrics and multiple model versions. It achieves strong accuracy, real-time efficiency, and improved handling of complex lane topologies, while ablations examine its conditional strategy, RIM, and transformer encoder.
- Experimental Setup: Experiments cover CurveLanes, CULane, and TuSimple, with F1 used for CurveLanes and CULane and FPR, FNR, and accuracy officially used for TuSimple.All datasets use 800×320 input images; RIM is applied only to CurveLanes because CULane and TuSimple lack fork-line cases.
- Qualitative Results: Visualization results show that CondLaneNet discriminates lane instances successfully in dense and fork-line cases.Different lane instances are represented by different colors across CurveLanes, CULane, and TuSimple visualizations.
- CurveLanes Results: 86.10 F1 on CurveLanes is 4.63% higher than CurveLane-L, while the small model reaches 85.09 F1 and 3.40% higher than SOTA.CurveLanes includes curve, fork, and dense lane cases; the authors associate the improvement with handling these complex topologies.
- CULane Results: 79.48 F1 on CULane is 3.19% higher, with best performance in eight of nine scenarios; the small model achieves 78.14 F1 at 220 FPS.The small model is reported to have similar efficiency to LaneATT-S while improving F1 by 4.01%.
- Ablation Studies: Ablations show gains from the conditional lane strategy, offset map, transformer encoder, and RIM, with RIM improving accuracy for fork and dense lines.Removing the transformer encoder causes a significant drop for the small model, while using the standard model’s proposal heatmap brings results close to the full model.
5. Conclusion
CondLaneNet is a top-to-down lane detection framework that detects lane instances before predicting each instance’s shape, with RIM addressing dense and forked topologies. It achieves state-of-the-art results while retaining real-time efficiency.
- CondLaneNet detects lane instances first and then predicts their shapes instance-wise.Its conditional lane detection strategy combines conditional convolution with a row-wise formulation.
- RIM addresses complex lane-line topologies such as dense lines and fork lines.
- CondLaneNet refreshes state-of-the-art performance on CULane, CurveLanes, and TuSimple.
- The small CondLaneNet surpasses other methods in accuracy while providing real-time efficiency on CULane and CurveLanes.